gpt-6-astra
Meant to excel at
-
Computer use for forms, CRM updates, web research and data analysis; OSWorld 2.0 72.6 percent in about 40 minutes.
Vendor's own site openai.com -
Produces slides, spreadsheets and documents that follow existing templates and a user's style.
Vendor's own site openai.com -
OpenAI's strongest coding model, asks clarifying questions only when answers matter, and handles new instructions mid-task.
Vendor's own site openai.com -
Top results on hard science and math: GPQA Diamond 96.0 and FrontierMath Tier 4 97.6 percent.
Vendor's own site openai.com -
Strong at 3D and CAD work; BenchCAD 95.9 percent and a Blender house built in about 40 minutes.
Independent evaluator hierarchy.ch -
Token-efficient with parallel swarm mode up to 40 subagents; one reviewer saw about 4x more work per dollar.
Independent evaluator finance.biggo.com -
First model to score 43/43 on Plotly's analytics test, finding all five hidden-data issues.
Practitioner report plotly.com -
Highest functional and secure-code pass rates among OpenAI models on Endor Labs' Codex benchmark (82.1% functional, 34.6% secure), with zero confirmed cheating (vendor-run benchmark).
Independent evaluator endorlabs.com
Watch out for
-
Quality swings between peaks and failures; it ignored explicit instructions, made unwanted changes and merged an unfinished PR.
Practitioner report finance.biggo.com -
Trails Fable 5.1 on coding finish and UI polish; may stop early or leave features looking done but broken.
Practitioner report hierarchy.ch -
Extra safety checks can halt legitimate work, including defensive security; advanced exploit tasks are refused.
Vendor's own site openai.com -
Ranked tenth on Vals AI's Excel modeling benchmark, weaker than Claude models on LBO and DCF tasks.
Independent evaluator vals.ai -
Highest cost on that test (about $1,161 per person per year estimate) with Sol matching it for about a tenth.
Practitioner report plotly.com -
Cost: about $2.34 per task on Endor's Codex benchmark vs $0.52 for GPT-6 Sol at only modestly lower security scores.
Independent evaluator endorlabs.com
Benchmarks
Coding percentile 0.96
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| lmarena code arena webdev | 1777.2 | 0.96 | Independent evaluator lmarena.ai | undated |
Agentic coding percentile 0.55
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| artificial analysis coding agent index | 62 | 0.20 | Independent evaluator artificialanalysis.ai | 2026-09-09 |
| deepswe v1.1 | 74.1 | 0.70 | Vendor's own site openai.com | 2026-09-03 |
| terminal-bench 2.1 (aggregator, anotherwrapper.com) | 87.3 | 0.50 | Independent evaluator anotherwrapper.com | 2026-10-19 |
| terminal-bench 4.0 | 59 | 0.81 | Independent evaluator artificialanalysis.ai | 2026-09-09 |
Knowledge percentile 0.69
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| humanity's last exam | 57.2 | 0.69 | Vendor's own site openai.com | 2026-09-03 |
Overall percentile 0.70
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| artificial analysis intelligence index | 53 | 0.70 | Independent evaluator artificialanalysis.ai | 2026-09-09 |
Reasoning percentile 1.00
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| gpqa diamond | 96 | 1.00 | Vendor's own site openai.com | 2026-09-03 |
Tool use percentile 1.00
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| automationbench-aa | 69 | 1.00 | Independent evaluator artificialanalysis.ai | 2026-09-09 |
Percentile is the share of tracked models this model beats on that benchmark, averaged across the group. Only benchmarks at least four tracked models share are scored.
Practitioner reports
- +
Astra was the first model to ace Plotly's 43-question analytics test (43/43) but was the costliest. Single run, private questions, one harness.
Task: data-analysis | Compared with claude-opus-5.5 | 2026-10-14
Practitioner report plotly.com - +/-
For one creator's business deliverables (deck, financials, dashboard), Astra produced functional but plain output, with a 17-slide deck and mostly basic sum formulas in Sheets. The post reports no Opus result for this task. Subjective, single creator, vendor blog.
Task: data-analysis | 2026-09-24
Practitioner report mindstudio.ai - -
Found Astra overkill as an orchestrator agent: it wasted most of its effort getting sidetracked. Also said it burns credits heavily (a second user, Supranova, made the same credit complaint on 2026-09-29).
Task: agentic-long-running | 2026-09-23
Practitioner report community.openai.com - +
Astra was a stronger editor and handled timed tasks better than Opus 5.5, though Opus was ahead on a latency-budget coding task.
Compared with claude-opus-5.5 | 2026-09-22
Practitioner report every.to - +
Astra won most head-to-heads with Fable 5.1 on coding and reasoning because it understood intent, but still needed oversight.
Compared with claude-fable-5.1 | 2026-09-13
Practitioner report mindstudio.ai - +/-
Gave Astra a roughly 3-hour goal-mode refactor (split dashboard into its own repo plus new design system). It handled implementation, debugging and verification with little supervision but the first result dropped the author's favourite feature; reaching a verified result took about 12h40 elapsed and exhausted his weekly Pro limit three times. Author thinks UI work may not be its strength and that it is only worth it for complex problems or big budgets.
Task: agentic-long-running | 2026-09-08
Practitioner report blog.vibecodingdad.com - +
Argues Astra suits an escalation role: repository-wide changes, terminal-heavy agents, SRE/debugging, long investigations and multi-tool workflows; not worth the price for classification, short rewriting or routine summaries. Post says it is adapted from a gateway vendor's research, so treat as opinion with commercial interest.
Task: agentic-long-running | Compared with gpt-5.6-sol | 2026-09-08
Practitioner report dev.to - +/-
Astra rebuilt a house from a floor plan in Blender and tested its own Unreal export, but stopped early on coding and left features unfinished.
Compared with claude-fable-5.1 | 2026-09-08
Practitioner report hierarchy.ch - +/-
Reviewer preferred Fable 5.1 for reliable production code but liked Astra's token efficiency, computer use and 3D rendering; Astra ignored instructions and swung in quality.
Compared with claude-fable-5.1 | 2026-09-07
Practitioner report finance.biggo.com - +/-
Recommends Astra at medium effort by default: citing a dev.to coding test, medium was cheaper and faster than Sol at high, while Astra at high was slower and its review pass missed a startup bug that the medium run caught. Suggests medium for diff review and extraction, high for long autonomous tool-call runs. Secondhand result.
Task: code-review | Compared with gpt-5.6-sol | 2026-09-06
Practitioner report ilikekillnerds.com - -
On one ordinary coding task in an existing codebase Astra used about 30% of a weekly limit on a $100 plan, reasoned carefully but was not dramatically better than GPT-5.6 Sol and still introduced its own bugs; author keeps Sol as default for 80-90% of coding and escalates to Astra only for large-codebase redesigns or hard bugs.
Task: feature | Compared with gpt-5.6-sol | 2026-09-05
Practitioner report bilimgram.medium.com - +/-
Astra's app clones were poor and chess play was blunder after blunder, yet computer-use work like book recommendations and medical-record review was strong.
Compared with claude-fable-5.1 | 2026-09-05
Practitioner report medium.com - -
Forum user reports refusals on cyber requests rose from roughly half to about 94% with Astra, a regression for security research work.
Task: research | 2026-09-03
Practitioner report community.openai.com - +
An Astra reviewer settled in zero to two review loops versus five to seven with other OpenAI models, and was less nitpicky than Fable.
Compared with claude-fable-5.1 undated
Practitioner report news.ycombinator.com